A Hybrid XGBoost-LSTM Approach to Day-Ahead Solar Forecasting across Diverse Climatic Zones: A Comparative Study of Six Nigerian Cities

Abstract

Accurate day-ahead solar forecasting is essential for integrating photovoltaic generation into Nigeria's power grid. This paper presents the development and evaluation of a hybrid Long Short-Term Memory (LSTM) – Extreme Gradient Boost (XGBoost) model for solar irradiance forecasting across selected Nigerian cities. The study utilizes satellite-based meteorological and solar irradiation datasets obtained from NASA POWER and European commission PVGIS covering the period from 2014-2024. The datasets included Global Horizontal irradiance (GHI), temperature, relative humidity, and wind parameters. The methodology involved data acquisition, preprocessing, normalization, feature engineering, and model training using google Colab. The LSTM network was deployed to capture temporal dependencies and sequential characteristics in solar irradiance data, while the XGBoost algorithm was utilized to model nonlinear relationships among meteorological variables. The outputs of both models were integrated using a weighted average hybrid approach to improve forecasting performance, with the model performance evaluated using the standard metrics. The results showed that the hybrid model outperformed the standalone LSTM and XGBoost models across all selected cities with MAE values between 0.027 and 0.042 across the six cities and R-squared exceeding 0.95 for the four inland locations (Abuja, Enugu, Kano, Maiduguri) and dropping to 0.916 for Lagos and 0.878 for Port Harcourt. XGBoost carried most of the predictive weight (alpha = 0.85–0.95), though the LSTM contributed more in coastal climates where weather variability benefits from temporal memory. The Hybrid model achieved lower forecasting errors and higher prediction accuracy, particularly in northern regions where solar irradiance patterns are more stable. While, Coastal cities recorded relatively higher prediction errors due to cloud variability and cloud cover. The methodology is readily transferable to other data-scarce regions in Sub-Saharan Africa and provides a foundation for grid operators to optimize dispatch scheduling, reduce fossil fuel backup dependence, and accelerate solar integration. The findings of this study demonstrate the proposed framework can significantly enhance solar irradiance forecasting accuracy in Nigeria. The developed model can support solar power generation planning, smart grid management, renewable energy integration, and energy policy development. this research contributes to renewable energy forecasting literature by providing a multi-city hybrid forecasting framework suitable for tropical and developing regions with diverse climatic conditions.